Su Ufku

8 forecasts have a result, and 7 landed inside the range.

We keep every forecast and check it against the reading when its date comes. Forecasts made on the day and tests run afterwards on past years are kept apart.

The model was corrected on the morning of 28 October 2026 (version 2). Forecasts from earlier mornings were made with the old model and stay on the scorecard as they were. The old model used the pipeline totals İSKİ reports, while it had learned that water from past data as a lower net level, so every forecast drifted up a little each day. The new version uses the level it learned. The main reason the short runs after the October rain came out too high remains: after a big rain the model spreads the water’s arrival in the dams over weeks, while the readings showed it arrived within a few days.

The short runs’ results are here too.

8 forecasts have a result, and 7 landed inside the range. With so few, we don’t give a hit rate.

Every day the model works out each day ahead from 1 to 90, and keeps them all. The short ones get a result first. They are not published forecasts and are kept out of any hit rate, but when they miss, they show up here too.

Table of the 120 short runs
Target dateDays aheadForecast and rangeWhat happenedResultModel version
9 November 2026163–64%61.9%outside the range2
9 November 2026264–65%61.9%outside the range2
9 November 2026364–65%61.9%outside the range2
9 November 2026462–64%61.9%outside the range2
9 November 2026562–67%61.9%outside the range2
9 November 2026661–68%61.9%inside the range2
9 November 2026761–67%61.9%inside the range2
9 November 2026860–67%61.9%inside the range2
9 November 2026959–67%61.9%inside the range2
9 November 20261059–67%61.9%inside the range2
9 November 20261159–71%61.9%inside the range2
9 November 20261261–75%61.9%inside the range2
9 November 20261362–84%61.9%inside the range1
8 November 2026164–65%62.0%outside the range2
8 November 2026264–65%62.0%outside the range2
8 November 2026362–63%62.0%inside the range2
8 November 2026462–67%62.0%outside the range2
8 November 2026561–67%62.0%inside the range2
8 November 2026661–67%62.0%inside the range2
8 November 2026760–66%62.0%inside the range2
8 November 2026859–67%62.0%inside the range2
8 November 2026959–66%62.0%inside the range2
8 November 20261059–69%62.0%inside the range2
8 November 20261160–74%62.0%inside the range2
8 November 20261261–82%62.0%inside the range1
8 November 20261363–86%62.0%outside the range1
7 November 2026163–63%61.9%outside the range2
7 November 2026261–62%61.9%outside the range2
7 November 2026361–65%61.9%inside the range2
7 November 2026461–66%61.9%inside the range2
7 November 2026561–66%61.9%inside the range2
7 November 2026659–65%61.9%inside the range2
7 November 2026759–65%61.9%inside the range2
7 November 2026858–64%61.9%inside the range2
7 November 2026959–68%61.9%inside the range2
7 November 20261060–72%61.9%inside the range2
7 November 20261161–81%61.9%inside the range1
7 November 20261263–85%61.9%outside the range1
7 November 20261357–88%61.9%inside the range1
6 November 2026159–60%61.5%outside the range2
6 November 2026260–64%61.5%inside the range2
6 November 2026360–64%61.5%inside the range2
6 November 2026460–64%61.5%inside the range2
6 November 2026559–63%61.5%inside the range2
6 November 2026658–64%61.5%inside the range2
6 November 2026758–63%61.5%inside the range2
6 November 2026859–66%61.5%inside the range2
6 November 2026959–70%61.5%inside the range2
6 November 20261061–78%61.5%inside the range1
6 November 20261163–82%61.5%outside the range1
6 November 20261257–86%61.5%inside the range1
6 November 20261356–79%61.5%inside the range1
5 November 2026159–61%59.4%inside the range2
5 November 2026258–62%59.4%inside the range2
5 November 2026359–62%59.4%inside the range2
5 November 2026458–62%59.4%inside the range2
5 November 2026558–62%59.4%inside the range2
5 November 2026658–62%59.4%inside the range2
5 November 2026758–64%59.4%inside the range2
5 November 2026859–68%59.4%inside the range2
5 November 2026961–75%59.4%outside the range1
5 November 20261062–79%59.4%outside the range1
5 November 20261157–84%59.4%inside the range1
5 November 20261255–77%59.4%inside the range1
5 November 20261355–76%59.4%inside the range1
4 November 2026157–59%54.8%outside the range2
4 November 2026257–60%54.8%outside the range2
4 November 2026357–60%54.8%outside the range2
4 November 2026457–60%54.8%outside the range2
4 November 2026557–60%54.8%outside the range2
4 November 2026658–62%54.8%outside the range2
4 November 2026759–66%54.8%outside the range2
4 November 2026860–72%54.8%outside the range1
4 November 2026961–77%54.8%outside the range1
4 November 20261057–81%54.8%outside the range1
4 November 20261155–75%54.8%outside the range1
4 November 20261255–73%54.8%outside the range1
4 November 20261355–70%54.8%outside the range1
3 November 2026156–57%54.2%outside the range2
3 November 2026256–57%54.2%outside the range2
3 November 2026356–58%54.2%outside the range2
3 November 2026456–58%54.2%outside the range2
3 November 2026557–60%54.2%outside the range2
3 November 2026658–63%54.2%outside the range2
3 November 2026759–67%54.2%outside the range1
3 November 2026860–71%54.2%outside the range1
3 November 2026956–78%54.2%outside the range1
3 November 20261055–73%54.2%outside the range1
3 November 20261155–71%54.2%outside the range1
3 November 20261255–69%54.2%outside the range1
3 November 20261354–68%54.2%inside the range1
2 November 2026155–56%53.7%outside the range2
2 November 2026255–57%53.7%outside the range2
2 November 2026355–56%53.7%outside the range2
2 November 2026456–58%53.7%outside the range2
2 November 2026557–61%53.7%outside the range2
2 November 2026658–64%53.7%outside the range1
2 November 2026760–69%53.7%outside the range1
2 November 2026856–73%53.7%outside the range1
2 November 2026955–71%53.7%outside the range1
2 November 20261055–69%53.7%outside the range1
2 November 20261154–67%53.7%outside the range1
2 November 20261254–67%53.7%outside the range1
2 November 20261354–68%53.7%inside the range1
1 November 2026155–56%53.7%outside the range2
1 November 2026255–55%53.7%outside the range2
1 November 2026356–57%53.7%outside the range2
1 November 2026457–60%53.7%outside the range2
1 November 2026557–62%53.7%outside the range1
1 November 2026659–66%53.7%outside the range1
1 November 2026756–70%53.7%outside the range1
1 November 2026855–68%53.7%outside the range1
1 November 2026955–66%53.7%outside the range1
1 November 20261054–64%53.7%outside the range1
1 November 20261154–66%53.7%outside the range1
1 November 20261254–66%53.7%inside the range1
1 November 20261353–67%53.7%inside the range1
31 October 2026154–55%53.8%outside the range2
31 October 2026255–56%53.8%outside the range2
31 October 2026356–59%53.8%outside the range2

The race against simple rules

We set the forecasts made on the same mornings, for the same target dates, beside the simple rules’ forecasts. Error is the average distance of the forecast from the measured level, in percentage points; smaller is better. “Closest” counts how often a method gave the forecast nearest the measurement.

14-day forecasts

Over 8 forecasts with a result, all from the same mornings.

MethodMean error (points)Closest on
Water-balance model (the published forecast)4.174 of 8 forecasts
No change (today’s level)5.654 of 8 forecasts
Same day last year9.080 of 8 forecasts
The usual change for this season7.870 of 8 forecasts
This date’s long-run average9.970 of 8 forecasts

30-day forecasts

None has a result yet. The first result is due on 18 November 2026.

1–13-day short runs

We don’t publish these; the model makes them every day and we keep them.

Over 195 forecasts with a result, all from the same mornings.

MethodMean error (points)Closest on
Water-balance model (the published forecast)3.2856 of 195 forecasts
No change (today’s level)2.73131 of 195 forecasts
Same day last year4.196 of 195 forecasts
The usual change for this season3.732 of 195 forecasts
This date’s long-run average6.960 of 195 forecasts

We make the forecast again every morning.

The live archive started on 19 October 2026. It holds 22 mornings’ forecasts so far. 7 days ago the most likely level a month out was 75%. In today’s forecast it is 83%. The forecast has moved up 8 points.

0%25%50%75%100%measured52–81%71–96%18 Oct25 Oct1 Nov8 Nov0%25%50%75%100%measured52–81%71–96%18 Oct25 Oct1 Nov8 Nov
Each upright line is one morning’s forecast: the range we gave for 30 days later, with a short tick at the most likely level. The dark area is the measured level. Once a forecast’s date arrives, what happened is added as a filled triangle. The thinner line beside each morning is that morning’s 14-day forecast. The thinnest line is the 7-day run, which we don’t publish.

Look at a past morning

Choose a morning to see the forecast we made that day, the level measured since, and how the forecast turned out.

The forecast we made on the morning of 19 October 2026

The level that morning: 52.4% (reading of 18 October 2026).

  • In 14 days, 2 November 2026: range 53–71%, most likely 58%. Measured 53.7%, inside the range.
  • In 30 days, 18 November 2026: range 52–81%, most likely 62%. The result is due on 18 November 2026.
40%50%60%70%80%forecast day11 Oct18 Oct25 Oct1 Nov8 Nov15 Nov

The model version that morning: 1.

A bar set in advance

Water-balance model (the published forecast)

3.7 points

Same day last year

5.7 points

Mean error of the 30-day forecast in the hindcast, in percentage points. The shorter the bar, the better.

The test is the simple rule “the level will be what it was on this date last year”. At 30 days the model is off by 3.7 points on average, the rule by 5.7. The model beats the rule, so the 60–90-day scenarios are shown.

Live scorecard

Forecasts made on the day, before the outcome was known.

The live scorecard started on 19 October 2026. In 22 days, 44 14- and 30-day forecasts were saved. 8 of them have a result, and 36 are waiting for their target date. Only the forecasts we publish, at 14 and 30 days, are counted here. Days 60 and 90 are scenarios, so they are not listed as a single forecast.

Live results by days ahead

Days aheadMean error (points)Share inside the rangen
14 days4.288%8

Saved forecasts and what happened, newest target date first

The dashed line is the forecast’s range and the short tick the most likely level. When the target date arrives, what happened is added as a thick upright mark.

  1. 9 December 2026made on 9 November 2026, 30 days ahead

    71–96% 83%

    not known yet

    waiting

  2. 8 December 2026made on 8 November 2026, 30 days ahead

    68–94% 80%

    not known yet

    waiting

  3. 7 December 2026made on 7 November 2026, 30 days ahead

    66–92% 79%

    not known yet

    waiting

  4. 6 December 2026made on 6 November 2026, 30 days ahead

    64–91% 76%

    not known yet

    waiting

  5. 5 December 2026made on 5 November 2026, 30 days ahead

    63–90% 76%

    not known yet

    waiting

  6. 4 December 2026made on 4 November 2026, 30 days ahead

    64–93% 77%

    not known yet

    waiting

  7. 3 December 2026made on 3 November 2026, 30 days ahead

    64–94% 78%

    not known yet

    waiting

  8. 2 December 2026made on 2 November 2026, 30 days ahead

    63–91% 75%

    not known yet

    waiting

  9. 1 December 2026made on 1 November 2026, 30 days ahead

    61–90% 74%

    not known yet

    waiting

  10. 30 November 2026made on 31 October 2026, 30 days ahead

    59–86% 70%

    not known yet

    waiting

  11. 29 November 2026made on 30 October 2026, 30 days ahead

    58–85% 68%

    not known yet

    waiting

  12. 28 November 2026made on 29 October 2026, 30 days ahead

    58–87% 69%

    not known yet

    waiting

  13. 27 November 2026made on 28 October 2026, 30 days ahead

    62–90% 74%

    not known yet

    waiting

  14. 26 November 2026made on 27 October 2026, 30 days ahead

    65–97% 80%

    not known yet

    waiting

  15. 25 November 2026made on 26 October 2026, 30 days ahead

    65–98% 80%

    not known yet

    waiting

  16. 24 November 2026made on 25 October 2026, 30 days ahead

    58–95% 72%

    not known yet

    waiting

  17. 23 November 2026made on 9 November 2026, 14 days ahead

    69–81% 74%

    not known yet

    waiting

  18. 23 November 2026made on 24 October 2026, 30 days ahead

    56–92% 71%

    not known yet

    waiting

  19. 22 November 2026made on 8 November 2026, 14 days ahead

    66–80% 71%

    not known yet

    waiting

  20. 22 November 2026made on 23 October 2026, 30 days ahead

    55–89% 67%

    not known yet

    waiting

  21. 21 November 2026made on 7 November 2026, 14 days ahead

    66–78% 70%

    not known yet

    waiting

  22. 21 November 2026made on 22 October 2026, 30 days ahead

    54–82% 64%

    not known yet

    waiting

  23. 20 November 2026made on 6 November 2026, 14 days ahead

    64–77% 68%

    not known yet

    waiting

  24. 20 November 2026made on 21 October 2026, 30 days ahead

    53–83% 65%

    not known yet

    waiting

  25. 19 November 2026made on 5 November 2026, 14 days ahead

    63–75% 68%

    not known yet

    waiting

  26. 19 November 2026made on 20 October 2026, 30 days ahead

    52–80% 62%

    not known yet

    waiting

  27. 18 November 2026made on 4 November 2026, 14 days ahead

    64–79% 69%

    not known yet

    waiting

  28. 18 November 2026made on 19 October 2026, 30 days ahead

    52–81% 62%

    not known yet

    waiting

  29. 17 November 2026made on 3 November 2026, 14 days ahead

    62–83% 69%

    not known yet

    waiting

  30. 16 November 2026made on 2 November 2026, 14 days ahead

    62–76% 68%

    not known yet

    waiting

  31. 15 November 2026made on 1 November 2026, 14 days ahead

    61–75% 67%

    not known yet

    waiting

  32. 14 November 2026made on 31 October 2026, 14 days ahead

    59–71% 64%

    not known yet

    waiting

  33. 13 November 2026made on 30 October 2026, 14 days ahead

    59–71% 63%

    not known yet

    waiting

  34. 12 November 2026made on 29 October 2026, 14 days ahead

    59–74% 65%

    not known yet

    waiting

  35. 11 November 2026made on 28 October 2026, 14 days ahead

    62–78% 69%

    not known yet

    waiting

  36. 10 November 2026made on 27 October 2026, 14 days ahead

    62–86% 73%

    not known yet

    waiting

  37. 9 November 2026made on 26 October 2026, 14 days ahead

    63–87% 73%

    61.9%

    outside the range

  38. 8 November 2026made on 25 October 2026, 14 days ahead

    57–89% 66%

    62.0%

    inside the range

  39. 7 November 2026made on 24 October 2026, 14 days ahead

    56–80% 66%

    61.9%

    inside the range

  40. 6 November 2026made on 23 October 2026, 14 days ahead

    55–78% 61%

    61.5%

    inside the range

  41. 5 November 2026made on 22 October 2026, 14 days ahead

    55–71% 60%

    59.4%

    inside the range

  42. 4 November 2026made on 21 October 2026, 14 days ahead

    54–70% 61%

    54.8%

    inside the range

  43. 3 November 2026made on 20 October 2026, 14 days ahead

    54–69% 58%

    54.2%

    inside the range

  44. 2 November 2026made on 19 October 2026, 14 days ahead

    53–71% 58%

    53.7%

    inside the range

All 44 rows are shown.

Hindcast

Forecasts made afterwards for past years, each time with only the data known on that day.

For 2011–2023 we re-made a forecast every week using only the data known on that day (679 start dates in all). No forecast used data from after its own date. The test shows how well the method would have done had it been used in those years.

In the hindcast the first 14 days were also replayed with past years’ weather. Live forecasts use a real weather forecast for those days. So at short range the hindcast may understate the real skill.

The model and the simple rules

Mean absolute error in percentage points: saying 53% when the dams are 50% full is an error of 3 points. The shorter the bar, the better. The simple rules we compare against: “no change”, “same day last year”, “this date’s long-run average” and “the usual change for this season”.

Method14 days30 days60 days90 days
Water-balance model (the published forecast)2.03.76.58.8
The usual change for this season2.24.17.09.2
Same day last year3.15.79.712.9
No change (today’s level)3.57.113.418.6
This date’s long-run average13.413.413.313.4
Machine-learning model (LightGBM, tried and dropped)2.24.37.610.2

How often the range held

80% of outcomes should land inside the range. The closer the share in the table is to 80%, the better the range is set.

Method14 days30 days60 days90 days
Water-balance model (the published forecast)85%82%79%79%
The usual change for this season67%68%66%66%
Same day last year76%76%73%73%
No change (today’s level)74%74%73%74%
This date’s long-run average85%85%85%85%
Machine-learning model (LightGBM, tried and dropped)51%55%52%54%

Error by year

Mean error of the 14- and 30-day forecasts for each year (points).

024620112014201720202023Mean error (points)14 days30 days02462011201320152017201920212023Mean error (points)14 days30 days
The chart’s data as a table
Year14 days30 days
20111.83.2
20122.24.3
20131.63.3
20143.15.8
20151.93.3
20161.94.2
20172.01.9
20182.44.2
20191.32.6
20201.32.8
20212.24.3
20222.03.9
20231.84.1

By season

Mean error by the season the forecast was made in. A cell is marked where the model is no better than the strongest simple rule (the usual change for the season).

14 days30 days60 days90 days
Winter (Dec–Feb)3.54.25.87.07.89.58.610.2
Spring (Mar–May)1.51.52.62.4model is not better3.93.7model is not better4.74.6model is not better
Summer (Jun–Aug)0.80.81.51.63.33.56.36.4
Autumn (Sep–Nov)2.12.24.95.311.111.315.815.7model is not better

Low-reserve years and the rest

The years in the test period when the city total fell below 21%: 2014, 2020, 2021 and 2023. The test uses 21% as its line, so 2020, whose lowest day stayed just above 20%, is in this group too.

14 days30 days60 days90 days
Other years1.91.93.43.56.15.8model is not better8.07.5model is not better
Low-reserve years2.12.84.25.47.69.710.813.2

In each cell the large figure is the model’s error and the small one below is the error of the “usual change for the season” rule.

The method, and where the model is weak